Md. Shakhoyat Rahman Shujon, MD Jahid Hasan Jim, Md. Milon Islam, Md. Rezwanul Haque, Fakhri Karray
5 min
Abstract
We introduce PAST-TIDE, our stance detection system addressing both subtasks of the StanceNakba Shared Task at NakbaNLP@LREC-COLING 2026. The main idea is statement tuning. We redefine stance as cloze-style masked language modeling (MLM), letting a verbalizer map label words to stance categories through the pre-trained MLM head rather than appending a randomly initialized classification head. We complement this with prototypical contrastive learning, which uses learnable class prototypes for batch-size independent contrastive training, and topic-conditional layer normalization for cross-topic Arabic stance detection. PAST-TIDE achieves macro-F1 scores of 0.75 for Subtask A and 0.74 for Subtask B on the official leaderboard, indicating that minimal architectural additions to a pre-trained model can remain competitive in low-resource settings.
Alex: Even with the right prompts, the model can still drift—it might gradually lose track of what "support" looks like versus "opposition," especially when it's only seeing a handful of examples at a time. So they create three "prototypes," one for each stance category. Think of them as ideal, averaged-out examples of what support, opposition, and neutrality each look like. Every new sentence gets compared to these three anchors, and the model learns by asking: which anchor does this most closely resemble?
Sam: So the prototypes act as stable reference points that keep the model grounded, even when the training data is thin?
Alex: Exactly. Standard approaches compare items within a single batch of training data, which is unreliable if that batch is small—you might just happen to have a bad mix. Prototypes are always present regardless of batch size, so the model always has something solid to orient itself against.
Sam: They tested this on Arabic data covering multiple different topics. Did the model get confused between those topics—mixing up the language used for, say, refugees versus trade agreements?
Alex: That was a real concern. Different topics use different vocabulary and framing, so they added a technique called "topic-conditional layer normalization." In plain terms, it means the model generates a slightly different set of internal settings for each topic—like adjusting the lens on a camera depending on what you're photographing. This stops the model from blending the linguistic patterns of one debate into another.
Sam: So it's a topic-specific filter built into the model's internals. Did the authors flag any limitations to all of this?
Alex: They did. Because of limited computing resources, they couldn't run a fully controlled experiment that isolates exactly how much each individual component—the fill-in-the-blank framing, the prototypes, the topic filter—contributes on its own. Their results come from comparing different versions of the full system rather than cleanly separating each piece.
Sam: Right—if you change three things at once, it's hard to know which one is doing the heavy lifting.
Alex: That's a fair characterization of the limitation they acknowledge. What they can say is that the combined system performs competitively in low-resource settings, and the underlying logic—using what the model already knows rather than adding complexity—holds up across their tests.
Sam: So the practical takeaway is that when data is scarce, the smarter move is to ask better questions of the model you already have, rather than building a bigger, more complex one?
Alex: That's the implication the authors draw. And they suggest the logical next step is automating the selection of the best fill-in-the-blank words—rather than having researchers choose them manually—which could make these systems more effective and easier to deploy.
Sam: It's a useful reminder that sometimes a small change in how you frame a problem—shifting from "teach the model a new skill" to "ask it to use what it already knows"—can matter more than adding layers of new machinery.
Alex: And that's what makes this work worth paying attention to. By leaning on the knowledge already embedded in pre-trained models, they achieved competitive results with almost no new trainable parameters. It's a case where careful design compensated for a real data constraint.
Sam: A good example of working smarter with limited resources. Thanks for walking us through it, Alex.
Alex: My pleasure. Thanks for listening to ResearchPod.